Moss-Image: Text-to-Image Generation with Native Layout Control

MOSS-Image is a text-to-image model family that adds native, structured layout control to prompt-driven image generation. It includes a 5B-parameter backbone (MOSS-Image-Base) and a curated-data fine-tuned checkpoint (MOSS-Image-SFT), both accessible with Apache-2.0 weights. Prompts may be natural language or JSON that specifies canvas metadata, labeled items (text or objects), and pixel-aligned bounding boxes (format y_min, x_min, y_max, x_max in an abs_centralized coordinate space). The system supports long prompts up to 4K tokens, strong on-image text rendering, and both anime and photorealistic styles. Layout inputs enable controlled composition and editing; an example travel-poster JSON demonstrates how title and illustration boxes map to generated output.

Deployment uses a shared Diffusers pipeline with Python and Gradio frontends, recommended sampling (flow-matching Euler, ~50 steps, guidance_scale ~3.0, uncond_skip_layers 22), and options for COMFYUI integration. Quick-start instructions show loading via DiffusionPipeline.from_pretrained with custom_pipeline and device/precision settings, and a Prompt Enhancer tool converts sketches or coordinates into compact JSON for generation. Fine-tuning yields measurable improvements in rendered text clarity and element organization (whiteboard examples), and weights plus code are hosted on Hugging Face and ModelScope for experimentation and domain adaptation.